Deepseek-R1-0528 weights have dropped! Thanks to admirable research and releases from @tngtech and @xlr8harder, we have a straightforward path to free-speech-respecting SOTA with *no additional training*. First, some background: @tngtech published a method for reducing censorship in R1 by locating and blacklisting the top n experts implicated in refusals. .07% experts blacklisted eliminated over half of refusals, and the resulting model benches better, too. Virtuously, tng published their vLLM patch enabling this analysis. @tngtech also created the 'chimera' model by merging V3-0324 shared experts with R1's routed experts. Chimera benches like R1, but is more concise. If a merge like that works again, that would be a huge boon in and of itself, since R1-0528 is very loquacious. But aside from that, it showed how modular R1 weights are. @xlr8harder found (as part of his SpeechMap free speech eval) that chimera eliminates MANY refusals. More than Perplexity-R1 and Microsoft's effort. @xlr8harder has also virtuously published his multilingual question bank for SpeechMap. @xlr8harder verified that R1-Zero is almost completely uncensored (which we knew). So how might one operationalize this? 1. Spin up @tngtech's patch of vLLM for Mixture of Controllable Experts. 2. Run the SpeechMap prompts that R1-0528 refuses, and the most similar prompts it doesn't. 3. Rank the experts by censoriousness per @tngtech's Mixture of Controllable Experts paper. 4. Knock out the top n most censorious experts for increasing values of n until it stops helping or the model performance tanks. 5. For each knocked out expert, replace with the corresponding expert from R1-Zero, V3-Base, and/or combination of those -> evaluate. 6. IFF the model is still refusing prompts, try targeted abliteration. i.e., knock out the 'refusal direction', but only on the top-n experts implicated in those refusals. This is promising because abliteration works, but it usually tanks model performance. Abliterating only a couple handfuls of experts seems likely to greatly reduce the hit to model performance. Optionally: 7. Reproduce the chimera merge with our newly de-censored R1-0528-decensored routed experts + V3-0324 shared experts and evaluate.